PosNet: A CNN-based Transformer Model for UE Localization in NLOS Dominated Scenarios

Abstract

Wireless positioning will play a pivotal role in enabling emerging use-cases such as autonomous driving, remote surgery, and Industry 4.0. Most of these applications demand sub-centimeter positioning accuracy, which is challenging to achieve with existing 5G-Advanced networks, especially in non- line-of-sight (NLoS) dominated scenarios. However, AI/ML-based techniques have the potential to extract multi-path propagation features and robustly associate them with the 2D location of the user equipment (UE). This work proposes a CNN-based AI/ML model, called PosNet, which leverages a transformer backbone to focus the model’s attention on the CIR samples that contain most amount of multi-path propagation information. When evaluated using realistic channel models (InF-DH) under practical transmitter and receiver constraints, the proposed model achieves a positioning error of 5.4cm for more than 90% of users. This represents an improvement of approximately 225% over the best existing AI/ML-based model and about 12,200% over the best traditional signal processing techniques.